An Explainable AI Approach for Diabetes Risk Prediction Using Logistic Regression and NICE Counterfactual Explanations

The rapid development of Artificial Intelligence (AI) in healthcare has accelerated the adoption of machine learning models for predicting chronic diseases such as diabetes. However, limited model interpretability remains a challenge for clinical decision-making. This study proposes an explainable diabetes risk prediction model by integrating Logistic Regression with the Nearest Instance Counterfactual Explanation (NICE) using the Pima Indians Diabetes Dataset. The workflow includes data preprocessing, median imputation, feature standardization, hyperparameter tuning using GridSearchCV, and recall-oriented threshold optimization. The optimal threshold of 0.32 achieved an accuracy of 74.03%, precision of 58.97%, recall of 85.19%, F1-score of 69.70%, and an ROC-AUC of 0.8222 (95% CI: 0.6957–0.9655). Odds ratio analysis identified Glucose (OR = 2.5798) and BMI (OR = 1.5752) as the strongest predictors of diabetes risk. Furthermore, NICE with locked non-actionable features generated realistic and actionable counterfactual recommendations, achieving 100% validity and actionability. The proposed framework demonstrates the potential to support interpretable Clinical Decision Support Systems (CDSSs) for early diabetes risk assessment, although medical experts have not clinically validated its recommendations.